dsh-agentic-router

When multiple model tiers (flash, mid, pro) are available, choosing the right one for every task is tedious and error-prone. dsh-agentic-router inspects task type and complexity at the start of each turn and automatically routes the request to the most suitable tier, removing the guesswork.

Under the hood, four expert recommenders (deterministic rules, UCB bandit, LinUCB, k-NN experience pool) propose candidates in parallel, and an EXP3 meta-selector weights the final decision. After each turn, the plugin computes a reward from quality proxy signals such as tool failure rate, model retries, actual cost, and latency, then feeds it back into the bandit weights so routing improves with every use. Starting with v1.5.0, step-level routing upgrades the next step to pro after a heavy tool call (code writing, execution, sub-agent) and falls back to the baseline tier for lightweight steps, approximating flash-for-simple and pro-for-complex. Every decision and reward is persisted as JSONL files for full auditability, replay, and state recovery after restart.

Ideal for DSH users who have multiple model tiers configured and want zero-config automatic routing. The default shadow mode records and learns without switching, so there is zero risk to try. Once satisfied, a single conversational command flips it to active mode, and the setting persists across restarts.